Summary

Follow-up to #6318 (the "auto-detect the model an agent CLI actually ran" half). The detection itself already exists across the agent-CLI drivers; this PR closes the one remaining gap — it was untested for claude_code.

State of the auto-detect mechanism (verified):

• codex_cli — parse_model_from_banner reads the model: line from codex's stderr startup banner and stamps it into CompletionResponse.actual_model (landed in codex cli 需要获取实际使用的模型 #6134). Already covered by test_parse_model_from_banner_*.
• claude_code — stream() captures the resolved model from the stream-json system/init event; complete() reads it from the result JSON's model field (forward-compat — claude -p --output-format json does not emit it yet, so that path stays None and the kernel falls back to the requested id). Both paths were wired but had no test.
• Kernel metering honours it uniformly: model: result.actual_model.clone().unwrap_or_else(|| model.clone()) (messaging.rs, agent_execution.rs, tools_and_skills.rs), so a driver that resolves its own model attributes the real one.

Change

Test-only. Two regression tests in crates/librefang-llm-drivers/src/drivers/claude_code.rs:

• stream_json_init_event_carries_resolved_model — deserializes a representative stream-json system/init line and asserts ClaudeStreamEvent.model is parsed (the streaming agent path's detection source).
• json_output_surfaces_model_when_present_else_none — pins both shapes of the result JSON: absent model → None (today's --output-format json behaviour, kernel falls back to requested id), present model → surfaced (forward-compat for a future CLI version).

These lock the deserialization contract so a refactor that drops or renames the serde model field fails loudly instead of silently turning resolved-model detection back into a no-op.

Not changed (and why)

• gemini_cli stamps actual_model: None. Gemini CLI emits no model banner/JSON field and runs exactly the --model it is given (no divergence), so there is nothing to detect; None correctly falls back to the requested id. Its CLI calls report zero tokens (billing is on the user's Google account), so the recorded model string is display-only and not a cost-attribution input — stamping the prefix-stripped id would only drop the gemini-cli/ provenance for no benefit.

Verification

In the worktree:

• cargo test -p librefang-llm-drivers --lib model — 24 passed, 0 failed (includes the two new tests)
• cargo clippy -p librefang-llm-drivers --all-targets -- -D warnings — clean
• cargo fmt -p librefang-llm-drivers -- --check — clean

Refs #6318. The manual-management half landed in #6327.